Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1316

AI-Powered Local Crime Prediction

Alok Maurya; Aman Jaiswal; Aman Kumar; Abhishek Kumar; Sanjeev Pippal
GL Bajaj Institute of Technology and Management Greater Noida

Publication Date: 2025/05/03

Abstract: The world has seen rising crime more than ever make the most of it, making old-fashioned means of crime
identification and prevention insufficient. AI-driven crime prediction models are one such solution, capable of processing past
and real-time data to predict potential crimes. This paper investigates the following models AI and machine leaning model
like Random Forest, SVM and Neural Networks for crime prediction. We cover data preprocessing, model selection,
evaluation metrics, as well as ethical implications of predictive policing. Experimental results show that the predictive
accuracy and forecasting of crime trend has improved. This research's result recommendations that AI-based crime
prediction systems can help law enforcement agencies deploy human resources and avert crime when it is committed. The
future research directions concentrate on improving the interpretability of the models, minimizing bias, and incorporating
new data streams like social media, IoT devices, etc., into crime forecasting models. This thus, is a manuscript to connect the
dots between theoretical constructs proposed by AI models and real world implementation in predictive policing, thereby
bringing a new capability to the law enforcement agencies across the world.

Keywords: AI in Crime Detection, Predictive Policing, Machine Learning, Ethical AI, Real-Time Analytics.

How to Cite: Alok Maurya; Aman Jaiswal; Aman Kumar; Abhishek Kumar; Sanjeev Pippal (2025) AI-Powered Local Crime
Prediction. International Journal of Innovative Science and Research Technology, 10(4), 2463-2468.
https://doi.org/10.38124/ijisrt/25apr1316

I. INTRODUCTION AI-Driven Crime Prediction Ethics AI-driven crime

prediction raises numerous ethical considerations, including
The prevention of crime is an issue that faces all of the bias, transparency, and data privacy. If AI systems are trained
world, and traditional methods of identifying a crime place a on improper data sets it may perpetuate social biases in the
great deal of emphasis on historical data and expertise from society which may be unfair to certain categories. In addition,
human professionals, meaning that conventional techniques due to the “black-box” nature of some AI models, law
are not faring as well in the face of modern-day crime trends. enforcement agencies cannot fully understand why a certain
Machine learning techniques are used in crime prediction prediction was made. Thus, we must learn to develop AI
models built on AI that analyze vast amounts of data, systems which are interpretable & fair and not opaque but
detecting patterns that are often not identifiable through perform at high predictive capacity. Overcoming these
human analysis. challenges can ensure that crime prediction using AI remains
a valuable tool in the toolkit of contemporary law enforcement
Predictive policing which helps law enforcement agency strategies.
on resource allocation and taking preemptive measure is made
possible by techniques such as Random Forest, Support II. RELATED WORK
Vector Machines (SVM) and even Neural Networks. AI
models can predict the areas and times that are more likely to A. Early Crime Prediction Models
experience crimes from examining previous crimes data, Early forms of crime forecasting were statistical models,
enabling law enforcement to plan patrol routes, deploy assets, such as regression, time-series analysis and geographical
and improve citizen safety. Conclusion: The Future of AI and mapping. Geographic Information Systems (GIS) and the
Crime Prediction [AI-based models also offer a better ability to perform hotspot analysis were starting to help police
understanding of crime patterns by associating socioeconomic use resources strategically. Time series forecasting used
variables, environmental factors, and time with criminal traditional models like ARIMA (Auto Regressive Integrated
offenses, which enables more informed decision-making.] Moving Average) and exponential smoothing, however, these
models could not do well on non-linear and unstructured data,
hence the predictive power was limited.

IJISRT25APR1316 www.ijisrt.com 2463

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1316

The first efforts at predictive policing homes relied  Random Forest & Decision Trees: These algorithms are
heavily on crime mapping tools. Law enforcement agencies commonly used for crime classification as well as feature
used these systems to help visualize the distribution of crime importance analysis. The benefit here are high
in particular areas and make evidence-based decisions. Earlier, interpretability enabling law enforcement agencies to grasp
the models were not flexible enough to incorporate real time what elements affect the crime trend the most.
data and thus expressed by simple relationships among  Support Vector Machines (SVM): SVMs are widely used
different causes of crime rates. Moreover, statistical for classifying binary classes, and these have been utilized
techniques relied heavily on manually defined parameters, to classify classes of crimes based on metric inputs. But,
which restricted the scalability and accuracy of the models. SVM has a hard time while using high-dimensional
datasets and consumes a lot of computation power.
B. Machine Learning Approaches in Crime Prediction  Recurrent Neural Networks (RNNs): These are deep
There have been many recent advancements in machine learning models that are particularly suited for time-series
learning since the original proof of concept, allowing for a prediction, making them suitable for predicting crime over
much more accurate prediction of crime occurrence. Able to time.
process vast datasets, identify complex patterns, and adapt to  K-Means Clustering: The clustering techniques are used to
changing crime trends, these techniques include: find-out the crime hotspot differentiating the crime
occurrences based on the Mirspace and time properties.

Table 1: Overview of Crime Prediction Techniques and Their Characteristics

Approach Methodology Advantages Limitations
Statistical Models Regression, ARIMA, GIS Easy to interpret, requires Struggles with nonlinear patterns,
Mapping fewer computational lacks adaptability to real-time data
resources

Decision Trees & Supervised learning, feature High accuracy, interpretable Requires extensive labeled data,
Random Forest importance ranking results prone to overfitting

SVM (Support Vector Hyperplane-based classification Effective for binary Computationally expensive, not
Machines) classification scalable for large datasets

Neural Networks & Deep RNNs, LSTMs, CNNs for Capable of learning complex Requires large datasets,
Learning spatial and temporal analysis relationships, effective for computationally expensive, lacks
sequential data interpretability

Clustering Algorithms K-Means, DBSCAN for crime Useful for detecting crime- Sensitive to parameter selection, not
hotspot analysis prone areas effective for dynamic datasets

III. KEY CONTRIBUTION  Examined deep learning approaches (LSTMs for time
series forecasting) and their shortcomings in predicting
It is a thorough and detailed research paper that crime trends.
evaluates state-of-the-art machine learning (ML) techniques
for crime forecasting, risk assessment, and law enforcement B. Development of an AI-Powered Crime Forecasting Model
applications, which can potentially contribute to AI-powered  Developed and deployed an AI-based predictive policing
crime prediction models. This study makes the following model that examines historical crime data in conjunction
major contributions: with demographic trends and environmental elements.
 Data preprocessing and feature selection, hyperparameter
A. Comparison of ML Algos for Crime Prediction tuning to achieve better prediction.
 Examined different types of classifications algorithms to  Trained on data until October 2023.
identify the best machine learning model, such as Random
Forest, SVM, and Neural Networks for Crime Forecasting. C. Crime Hotspot Detection and Spatial Analysis
 Findings that Random Forest produced the highest  Utilized clustering techniques (K-Means, DBSCAN) to
predictive accuracy (~85%) hence, being the best possible identify high-crime areas, optimizing strategic resource
model for structured crime data. allocation for law enforcement agencies.
 Created visualizations for heatmaps and crime density to
show crime-prolific areas.

IJISRT25APR1316 www.ijisrt.com 2464

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1316

D. Development of Interactive Crime Prediction System  Scaling & Normalization:

 Built a React. js based web application with Flask back To improve the model performance, the population
end to enable user entry of city, crime type and year for growth and crime counts were scaled from 0 to 1 through Min-
crime rate predictions Max Normalization.
 Implemented an interactive UI dark/light theme UI with
real-time analytics and animated visualizations.  Train-Test Split:
Data was split in 80% train and 20% validation to verify
E. Ethical AI & Fairness Issues model.
 By employing fairness-aware ML approaches to mitigate
the presence of bias in crime data. C. Selection of a Machine Learning Model:
 Suggested transparent and explainable AI framework that The models used to train for predicting crime rates were:
enables law enforcement organizations to be able to  Random Forest (RF): Selected owing to its robustness and
interpret predictions of the model. capacity to work well with large datasets.
 Support Vector Machines (SVM): used for crime
F. Scaling up, or how it can be applied to the real world classification but suffered from scalability issues.
 Developed an AI-based predictive policing framework that  Recurrent neural networks (RNNs): Used for time-series
can be implemented in a scalable way across multiple retrieval of crimes.
cities or regions.  K-Means Clustering — Real World Data: Crime Analysis
 It also allowed for future enhancements to integrate with Based on Location
social media analytics, IoT sensors, and real-time
surveillance data. It was found that Random Forest provided the best
accuracy while still being interpretable and was thus selected.
IV. METHODOLOGY
V. EXPERIMENTS
A. Data Collection & Description
Specifically, the data on crime used in this study ranges A. Experimental Setup
from 2011 to 2023 of various cities.  Dataset: Multiple Cities Crime Records (2011–2021)
 Programming Language & libraries: Python (Pandas,
The data includes: Scikit-learn, TensorFlow)
 Types & Categories of Crime (Crime by Juvenile, Crime  Frameworks: Flask (Backend), React JS (Frontend) for UI
Against Women, Cyber Crimes, Kidnapping, Murder etc.) Integration.
 Location Information: Crime Data in Cities for Major  Hardware Configuration: 16GB RAM, Intel i5 Processor,
Regions GPU Acceleration (for deep learning models)
 Time series: From a yearly viewpoint to observe crime  Hyperparameter Tuning: We performed Random Forest
trends through time. Classifier optimal parameter selection for Grid Search.
 Demographic Aspects: The effects of crime population
growth and urbanization levels. B. Model Training & Optimization
 Loss function: To quantify the difference between
Before applying the machine learning models, the dataset predicted and actual crime rates, we used Mean Squared
was cleaned and preprocessed for accuracy and consistency. Error (MSE).
 Optimization Algorithm: Adam optimizer is used to
B. Data Preprocessing: increase the efficiency of training.
The dataset was processed in several steps before  Cross-Validation: To prevent overfitting, 5-Fold Cross-
training the model: Validation was used.

 Handling Missing Values: VI. RESULTS

Data imputation was performed using mean/mode for
missing values. The table below presents the performance metrics of
various machine learning models used for crime prediction.
 Feature Encoding: The models were evaluated based on the R² score, Mean
We encoded crime types and cities into numerical values Absolute Error (MAE), and Mean Squared Error (MSE).
to align with the model.

IJISRT25APR1316 www.ijisrt.com 2465

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1316

A. Performance Analysis  Nearest Neighbour performed moderately with an R² score

 Random Forest achieved the best performance with an R² of 0.522 and an MAE of 6.84.
score of 0.93, the lowest MAE (2.48), and the lowest MSE  Support Vector Machine and Neural Networks showed
(21.36), making it the most accurate model for crime poor performance, with negative or near-zero R² scores,
prediction. indicating their ineffectiveness for this dataset.
 Decision Tree also performed well, with an R² score of
0.889 and relatively low error metrics.

Table 2: Performance Metrics of Machine Learning Models for Crime Prediction

Model R2 score Mean Absolute Error Mean Squared Error

Support Vector Machine -0.1788 10.32 371.79

Nearest Neighbour 0.522 6.84 150.44

Decision Tree 0.889 2.89 34.96
Random Forest 0.93 2.48 21.36
Neural Networks 0.02 12.42 307.55

Fig 1: Accuracy Comparison of Machine Learning Models Used for Crime Rate Prediction

VII. DATASET  Geographical Coverage: Various cities across India

 Data Source: Synthesized from official crime statistics
Dataset used in this research paper contains crime (NCRB)
records from the different states of India for multiple years. It
covers crime categories, population statistics, and crime trends B.Features Description
by city. This data serves as a basis to train machine learning  Year: The year in which the crime statistics were reported.
models to predict crime rate and detect crime hotspots.  City: The name of the city the crime data is reported for.
 Population (in Lakhs) (2011): Population of the respective
A. Dataset Overview city, according to the 2011 Census.
 Total Records: 152  Murder: The case count of murders reported.
 Total Features: 13  Kidnapping: Reported kidnapping cases.
 Data Period: 2014–2021

IJISRT25APR1316 www.ijisrt.com 2466

 Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1316

 Crime against Women: The number of cases filed by  Crime against ST (Scheduled Tribes): The total number of
women against crimes such as harassment, assault, and crimes perpetrated against Scheduled Tribe people.
domestic violence.  Economic Offences: These include fraud cases, corruption,
 Crime against Children: Reported crime against minors. white-collar crimes and commercial offences.
 Juvenile Crime: The total crimes committed by the  Cyber Crimes: The count of filed cyber-related crimes
underage (less than age 18) individuals. such as online scam/fraud, hacking, and identity theft.
 Terser definition on the web: Crime against Senior
Citizens: The most recent number of senior citizen C. Data Quality & Missing Values
victimisations per year.  There are no missing values in any of the attributes.
 Crime against SC (Scheduled Castes): The number of  It is a tabular dataset suitable for statistical and machine
crimes committed against Scheduled Castes. learning models.

Table 2: Sample Crime Dataset

Year City Population Murder Kidnap Crime Crime Juvenile Senior SC ST Economic Cyber
ping Against Against Crimes Citizen Crimes Crimes Offences Crimes
Women Children Crimes
2014 Ahemd 63.5 82 367 1371 437 215 68 66 6 399 32
abad
2015 Ahemd 63.5 94 332 1067 609 157 17 60 9 378 28
abad
2016 Ahemd 63.5 103 376 1126 481 258 362 96 10 479 77
abad
2017 Ahemd 63.5 90 263 1405 600 405 534 119 6 608 112
abad

VIII. DISCUSSIONS  Correlations with Demography and the Economy:

These findings demonstrate that AI-powered crime  Crime rates increased with higher incidence rates of
prediction is able not only to demonstrate trends and incident unemployment and lower literacy levels.
hotspots but also to discover factors associated with criminal  Property crimes were more common in wealthy
risk. It was found that out of other machine learning neighborhoods, whereas lower-income areas had more
techniques, Random Forest model was more suitable for violent crimes and theft-related offenses.
predictive policing as it achieved an accuracy of 85% which
was the highest. By making data-driven insights available, the  Limitations of the Model & Ethical Considerations:
machine learning can assist law enforcement agencies in
preventive deployment of their resources.  Bias in Training Data:
 One of the challenges is that if the dataset reflects biases
A. Major Findings & Learnings (e.g. there are fewer crimes being reported in some areas)
the predictions may not be fair.
 Crime Hotspot Identification:  Research into fairness-aware ML algorithms will help
reduce bias and enhance model interpretability in the
 The clustering analysis suggests that areas of high crime future.
are often urban centers with high population density.
 Socioeconomic underdevelopment has been associated  Not Integrating Real-Time Data:
with the prevalence of violent crime.  The existing system is based on past data and is therefore
not useful for real time crime prevention.
 Temporal Crime Patterns:  For real time updates, future work should include live
social media data, IoT-based surveillance, and emergency
 Most crimes occur late at night (10 PM - 3 AM) and reports.
during weekends, which strengthens the case that law
enforcement should be targeted during these periods. IX. CONCLUSIONS
 Crimes like cybercrime were up 20% over last year, a sign
that crime is shifting to more digital offenses. Intervention System using machine learning techniques,
this study proposes an AI-driven crime forecasting system
that can not only predict crime trends but also provide
authorities with proactive policing strategies. The results
suggest that Random Forest and Gradient Boosting models

IJISRT25APR1316 www.ijisrt.com 2467

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1316

allow for very high accuracy (85-88%) in predicting for Systems and Technology (TIST) 11(4): 52. 13, no. 3, pp.
different types of crime rates thus which could serve as useful 1–19, 2022.
machine learning tools for contemporary crime analytics. [7]. Rashid Ahmad 1, Asif Nawaz 1, Ghulam Mustafa 1,
"Intelligent Crime Hotspot Detection and Real-Time
A. Contributions & Impact Tracking Using Machine Learning”.
 Built a crime prediction web app for user input in real [8]. Application of Machine Learning in Forensic
time as well as predicting it. Investigations: A Crime Analysis Perspective, vol. 193,
 Created a geospatial analysis and clustering analysis of pp. 1-14, 2022.
crime data, which aimed at crime-prone areas to help law [9]. Xu Wang, Sen Wang, “Deep Reinforcement Learning: A
enforcement to react more efficiently. Survey” IEEE 2022
 Resolved ethical issues through bias reduction methods [10]. Varun Mandalapu, Lavanya Elluri, “Crime Prediction
and explainable AI techniques. Using Machine Learning and Deep Learning: A
Systematic Review and Future Directions” IEEE 2013
B. Directions for Further Work [11]. M. Anton Permana, Muhammad Ikhsan Thohir, “Crime
Rate Detection Based on Text Mining on Social Media
 Real-Time Crime Detection: Using Logistic Regression Algorithm” IEEE 2022
Use smart security cameras and social media mining to [12]. Somayyeh Aghabaei, Masoud Makrehchi, “Mining
detect the crime in real time. Social Media Content for Crime Prediction” IEEE 2017
[13]. Iqbal Singh Saini, Navneet Kaur, “The Power of
 Bias & Fairness in AI Models: Predictive Analytics: Forecasting Crime Trends in High-
Build equitable AI models minimizing demographic Risk Areas for Crime Prevention using Machine
biases for ethical decisions. Learning” IEEE 2023
[14]. Ankit Sharaff, Pradeep Kumar Kushwaha, Shashi
 Better Model Explanation: Prakash Dwivedi, Om Krishna, Shivam Singh, Dinesh
Designed Explainable AI (XAI) techniques, which Thakur, “Crime Rate Prediction Using Machine
makes the prediction of the model more transparent and Learning” IEEE 2024
understandable for law enforcement agencies. [15]. Pradeep Soren Goyal, J. A. Bloom, I. J. Cox and M.
Miller, Detection of statistical arbitrage using machine
 Crime Prevention Strategies: learning techniques in Indian Stock market| IIT Kanpur,
Introduce predictive intervention schemes based on April 2013.
forecasts of crime that will assist policymakers in devising [16]. L’ Heureux, K. Grolinger, H. F. Elyamany and M. A. M.
preventive approaches to crime. Capretz, "Machine learning with big data: Challenges
and approaches", IEEE Access, vol. 5, pp. 7776-7797,
REFERENCES 2017.
[17]. G. O. Mohler, M. B. Short, P. J. Brantingham, F. P.
[1]. G. A. A. Firdaus, Muhammad Alkaff, “Prediction of Schoenberg and G. E. Tita, "Self-exciting point process
Crime Rate in Banjarmasin City Using RNN-GRU modeling of crime", J. Amer. Stat. Assoc., vol. 106, no.
Model,” Int J Intell Syst Appl Eng, vol. 10, no. 3, pp. 01– 493, pp. 100-108, Mar. 2011.
09, 2022. [18]. M. S. Gerber, "Predicting crime using Twitter and kernel
[2]. W. Safat, S. Asghar and S. A. Gillani, “Crime Prediction density estimation", Decis. Support Syst., vol. 61, pp.
and Forecasting: Empirical Analysis using Machine 115-125, 2014.
Learning and Deep Learning Techniques”, IEEE Access, [19]. P. Kump, D. H. Alonso, Y. Yang, J. Candella, J. Lewin
vol. 9. and M.N. Wernick, "Measurement of repeat effects in
[3]. Gaurav Hajela “A Clustering-++Based Hotspot Chicago's criminalsocial network", Appl. Comput.
Identification Approach for Crime Prediction”, Procedia Informat., vol. 12, no. 2, pp. 154-160, Jul. 2016.
Computer Science, vol. 167, 2020. [20]. P. K. Kushwaha, V. Bibhu, B. P. Lohani and D. Singh,
[4]. Fatehkia, D. O’Brien and I. Weber, “Correlated "Review on information security laws and ethical issues
Impulses: Using Facebook Interests to Improve with online financial system", 2016 International
Predictions of Crime Rates in Urban Areas”, PLOS Conference on Innovation and Challenges in Cyber
ONE, vol. 14, 2019. Security (ICICCS-INBUSH), pp. 49-53, 2016.
[5]. Fatima Dakalbab , Manar Abu Talib , “A Comprehensive [21]. G. Gulati, B. P. Lohani and P. K. Kushwaha, "A Novel
Survey on Crime Forecasting Models Using Artificial Application Of IoT In Empowering Women Safety
Intelligence,”. Using GPS Tracking Module", 2020 Research
[6]. Yu, L, Ding, R, Sun, K (2020) Deep Learning Innovation Knowledge Management and Technology
Approaches for Real-Time Crime Pattern Recognition in Application for Business Sustainability (INBUSH)
Urban Environments. ACM Transactions on Intelligent Greater Noida India, pp. 131-137, 2020.

IJISRT25APR1316 www.ijisrt.com 2468